Research On Road Obstacle Detection Algorithms Based On Disparity Space

Real time environment perception system is a key and fundamental task of autonomous driving technology.Depending on virous sensors,the system collects environmental information constantly around itself.After obtaining dense and sufficient 3D scene data,how to accurately describe the obstacle distribution in traffic scene has become a new research topic.Since the artificial road environment is occupied by horizontal ground and vertical planes,and these planes present individual pattern in 3D space which determine its characteristics in the disparity space.Due to this fact,this paper explores a generic obstacle detection algorithm based on disparity space.This paper adopts a compact intermediate representation called "Stixel" to segment the obstacle region in the image domain.According to the characteristics of the sensors,the first part compares the stereo vision,structured light based and time-of-flight based sensor.The binocular stereo camera is chosen as the depth information sensor.Firstly,an accurate binocular stereo vision hardware system was built.The camera parameters are calibrated by Zhang's calibration method.The Bouguet algorithm is used to rectify the pairs of binocular images.Aiming to obtain a fine disparity map as input to the subsequent detection algorithm,then an improved SGM algorithm for outdoor scenes is proposed.Furthermore,an obstacle detection algorithm based on disparity space is proposed.It is a bottom-up approach in the image domain,and multiple independent steps are cascaded to achieve “Stixel” masked region of interest.This novel representation is used to avoid the processing of a large number of original 3D point cloud data.At the same time,a motion estimation model is designed,which combines the dynamic programming algorithm to obtain dynamic obstacles.Finally,the improved stereo matching algorithm is tested on the Middlebury stereo vision dataset,and the mismatch rate is lower than the original algorithm.Based on the KITTI visual dataset,HCI benchmark dataset and 6d-vision dataset,the experimental results show that the accuracy is better than the original algorithm,and the video sequence can be processed in real time.The contribution of this paper is to prove that the method of using disparity space map instead of three-dimensional point cloud or voxel can detect obstacles effectively and reliably.Moreover,this paper expands the application of binocular stereo vision in driving scene analysis,and provides a fast and reliable road obstacle detection method.